Modelling Concession Across Speakers in Task-Oriented Dialogue
Kavita Elisheba Thomas, Colin Matheson · 2003
We determine criteria for modellingconcession signalled via plan-based“but” in task-oriented dialogue (TOD)(following (Thomas, 2003)) basedon some examples from the corpora,analysing where Speech Act (SA) andplanning information fails to predictconcession. In our approach we focuson cases involving cross-speakerconcession, where the speaker acceptspart of the previous speaker’s turn butsignals contrast via “but”, and we arguethat this contrast can (in the examplesshown) be modelled as concessive.Then given a representation of task-planhistory in the Information State (IS)model of the dialogue, we present aninitial framework for an algorithm thatpredicts concessive interpretation acrossspeakers in two situations encounteredin the corpora. We motivate this workby showing how it updates beliefs in thePTT (Poesio and Traum, 1998) modelof dialogue and can be used to facilitaterecognition of planning mismatches andpotentially avoid misunderstandings,and more generally improve discourseunderstanding and natural languagegeneration (NLG).